Govoice

Inspiration

Optimization problem in generic office is a real problem and by optimizing certain KPI can greatly gives benefit starting from reduced cost to productive speed. My goal is to create software that covers more market by generalizing my solution to common finance problem in companies that deals with invoices for purchasing.

What it does

You start by giving the platform your invoices. For this demo, MeDo have not yet include plugin to integrate with real email provider, but our original requirements is to be able to pull from inbox. Once you upload invoices, let the work run itself and depending on configuration, your invoices will be processed such like:

Data entry -> Rule conditions (seperate invoices based on condition e.g. due amount) -> AI analyze for critical summary (relationship with supplier, longest possble payment after due) -> Routes to manager inbox for approval.

The platform is seperate into Staff and Manager user. The manager sees inbox of approval request, the staff sees dashboard of tasks, also configuring the rules and policies.

How we built it

We tried MeDo by writing a complete markdown of requirements but minimum technical specs, in hope to let MeDo do its magic on giving a solution. But with this method, we ended up with a product with precise solution as what we asked, no more no less. Hence in choosing to long detailed prompt, we have to be details as writing a skill markdown so we get what we write.

Then we try the normal prompting way, iterations of chat to refine into what we need, and surprisingly by "complaining", MeDo understood what we asked and try to innovate by asking relevant question to refine the prompt

Challenges we ran into

Writing with detailed prompt require an actual detailed of requirements such as we seen in real work product requirements document. We tried wrote a long prompt without any technical specs, but MeDo completely satisfy all we asked without more, which is a problem because our mindset was by "wanting" more gives them cognitive reasoning to add more, yet it result in precise.

Hence with normal iterating way, MeDo gets to specify their own requirements and we refine it by prompting over and over again. It did well as we expected before, but with caveat such as subtle bugs and logic behavior that doesn't align, but did align in the business perspecitve matter.

For example, I wrote

Make the config panel looks like puzzle where each node adds more predicate to the rules

MeDo comes with complete nice interface of puzzle but did not integrate the technical requirement well with the code

Accomplishments that we're proud of

We refine our prompt based on many unsatisfying result beforehand, and we do see results as we refine prompt into more structured language that LLM mostly understand, not in a humanic way of complaining and wanting stuff in easy mode, because if it were other human, than that human would understand in emotional sense.

Plugins we use

MeDo provides us OCR and LLM agents, which is the core MVP of our product.

Built With

• llm
• medo
• ocr
• react